ViSaRL: Visual Reinforcement Learning Guided by Human Saliency